<style type="text/css"> .right-column{ padding-top: 0; } </style> .left-column[ ```r Missing The Point Non-Convergence In Iterative Imputation Algorithms Hanne I. Oberman, Stef van Buuren, & Gerko Vink ARTEMISS workshop ICML 2020 ``` ] .right-column[ # Algorithmic Convergence With iterative imputation, the validity of inferences relies on algorithmic convergence. Signs of non-convergence (i.e., non-mixing, trending) are typically identified through visual inspection. <br> <!-- .footer[Missing The Point | Oberman, van Buuren, Vink | 2020 | github.com/hanneoberman/MissingThePoint] -->
] --- .left-column[ ```r Missing The Point Non-Convergence In Iterative Imputation Algorithms Hanne I. Oberman, Stef van Buuren, & Gerko Vink ARTEMISS workshop ICML 2020 ``` ] .right-column[ # Simulation Study When imputing an incomplete multivariate normal set ( `\(n_{\rm obs}=\)` 1000, `\(n_{\rm sim}=\)` 1000), we obtain valid regression estimates after 5 to 10 iterations. <!-- Multivariate normal set ( `\(n = 1000\)` ), 'amputed' with a varying proportion of incomplete cases, and 'imputed' with a varying number of iterations. Depicted are the bias and coverage rate of an estimated regression coefficient ( `\(n_{\rm sim} = 1000\)` ). --> <br> .pull-left[
] .pull-right[
] ] --- .left-column[ ```r Missing The Point Non-Convergence In Iterative Imputation Algorithms Hanne I. Oberman, Stef van Buuren, & Gerko Vink ARTEMISS workshop ICML 2020 ``` ] .right-column[ # Simulation Study (2) Whereas non-convergence diagnostics (potential scale reduction factor, auto-correlation) identify signs of non-convergence up-to 30 to 50 iterations. <br> .pull-left[
] .pull-right[
] ] --- .left-column[ ```r Missing The Point Non-Convergence In Iterative Imputation Algorithms Hanne I. Oberman, Stef van Buuren, & Gerko Vink ARTEMISS workshop ICML 2020 ``` ] .right-column[ # Take-away We conclude that—in the cases considered—it never hurts to iterate longer, but such calculations hardly bring added value. Read more on [github.com/hanneoberman/MissingThePoint](https://github.com/hanneoberman/MissingThePoint). Or follow my updates through Twitter [@hioberman](https://twitter.com/hioberman). .pull-left[ <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> <br> Upcoming... ] .pull-right[ <img src="https://github.com/gerkovink/shinyMice/blob/edits/ICML/shinymicehex.png?raw=true" width="255"> ] ] <!-- .footnote[Slides created via the R package [**xaringan**](https://github.com/yihui/xaringan).] -->